[BUG] Cannot use self-hosted OpenAI-compatible API in offline/restricted environments; model list and AI chat mostly inaccessible

[991547436]

Problem (one or two sentences)

In completely offline or restricted network environments, Roo-Cline fails to work with self-hosted OpenAI-compatible endpoints. Observed issues include:

Model list cannot be loaded
AI chat functionality mostly inaccessible

Logs indicate DNS resolution failures or connection timeouts, causing the extension to fail initialization and become largely unusable.

Context (who is affected and when)

all

Reproduction steps

Environment
VSCode Version: 1.112.0
Extension Version: roo-cline 3.51.1
OS: Windows 10 x64
Network: Offline / restricted environment

Expected result

Model dropdown should load correctly regardless of whether a proxy/capture tool is active

Actual result

Model dropdown does not populate

Variations tried (optional)

In a fully offline or restricted network environment, the Roo-Cline extension fails to load the model list by default. However, if Fiddler (or any local proxy/capture tool) is running, the model dropdown loads normally.

Developer tools suggest that the issue may be caused by a race condition or incorrect initialization code: the code responsible for fetching and rendering the model list does not trigger when no proxy is running. This indicates a potential bug in the extension’s startup or network detection logic.

Initialization code should not depend on side effects of network interception tools

App Version

3.51.1 (7c9722b)

API Provider (optional)

OpenAI Compatible

Model Used (optional)

all

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

rror fetching models for roo: Error: Failed to fetch Roo Code Cloud models: Request timed out after 10 seconds.


 Error fetching Vercel AI Gateway models: {
  "code": "EAI_AGAIN",
  "message": "getaddrinfo EAI_AGAIN ai-gateway.vercel.sh",
  "stack": "Error: getaddrinfo EAI_AGAIN ai-gateway.vercel.sh\n\tat hM.from 


R [Extension Host] [getRooModels] Error fetching Roo Code Cloud models: {message: 'This operation was aborted', name: 'AbortError', stack: 'AbortError: This operation was aborted\n\tat node:in…nc.roo-cline-3.51.1\\dist\\extension.js:4827:21890)', url: 'https://api.roocode.com/proxy/v1/models', hasApiKey: false}

[roomote-v0]

I've analyzed this issue and here's my proposed implementation plan:

Root Cause Analysis

In offline/restricted network environments, several model fetchers hang indefinitely because they use axios.get without any timeout. This causes two cascading failures:

1. Backend requestRouterModels handler blocks for minutes: When the webview loads, it requests router models for all providers (openrouter, requesty, unbound, vercel-ai-gateway, roo). The handler uses Promise.allSettled which waits for ALL providers to resolve. Without timeouts, DNS resolution failures on external hosts (e.g. openrouter.ai, ai-gateway.vercel.sh, api.roocode.com) can hang for 30-120+ seconds each.

2. Webview-side 10-second timeout fires prematurely: useRouterModels.ts has a 10-second timeout. Since the backend is blocked waiting for DNS failures, the routerModels response never arrives in time, causing the query to fail.

3. OpenAI Compatible model list requires an API key: The requestOpenAiModels handler in webviewMessageHandler.ts (line 1175) guards with if (message?.values?.baseUrl && message?.values?.apiKey), but many self-hosted endpoints do not require an API key. The actual getOpenAiModels() function already handles missing API keys gracefully.

Why Fiddler fixes it: A local proxy intercepts DNS resolution and handles it locally, so lookups either succeed through the proxy or fail fast instead of hanging.

Proposed Changes

1. Add timeouts to all model fetchers missing them

The following fetchers use axios.get without a timeout option and will hang indefinitely on DNS failures:

• src/api/providers/fetchers/openrouter.ts - axios.get at line 102 (no timeout)
• src/api/providers/fetchers/requesty.ts - axios.get at line 21 (no timeout)
• src/api/providers/fetchers/unbound.ts - axios.get at line 17 (no timeout)
• src/api/providers/fetchers/vercel-ai-gateway.ts - axios.get at line 61 (no timeout)

Fix: Add { timeout: 10000 } (10 seconds) to each axios call, consistent with the existing timeout used by the roo fetcher. This ensures DNS failures fail fast rather than hanging indefinitely.

2. Allow requestOpenAiModels to work without an API key

In src/core/webview/webviewMessageHandler.ts line 1175, change the guard from:

if (message?.values?.baseUrl && message?.values?.apiKey) {

to:

if (message?.values?.baseUrl) {

The underlying getOpenAiModels() function already handles a missing API key correctly (it simply omits the Authorization header). This change allows self-hosted endpoints that do not require authentication to populate their model lists.

3. Add timeout to getOpenAiModels axios call

In src/api/providers/openai.ts line 564, the axios.get for fetching OpenAI-compatible model lists also lacks a timeout. Add { timeout: 10000 } for consistency and to prevent hangs on unreachable endpoints.

Summary

These changes ensure that:

• All external API calls fail fast (within 10 seconds) in offline environments instead of hanging indefinitely
• The OpenAI Compatible provider can fetch model lists from self-hosted endpoints that do not require API keys
• The extension remains fully usable with local/self-hosted providers even when external services are unreachable

Please mention @roomote in a comment saying that you approve this plan, and I'll proceed with the implementation.